Sunteți pe pagina 1din 10

RECENT ACHIEVEMENTS IN OFF-LINE HANDWRITING

RECOGNITION SYSTEMS
B. VERMA, M. BLUMENSTEIN & S. KULKARNI
School of Information Technology
Griffith University Gold Coast Campus
PMB 50, Gold Coast Mail Centre, Qld 9726 Australia
E-mail: {B.Verma, M. Blumenstein}@gu.edu.au

This paper reviews the current state of the art in handwriting recognition research. The
paper deals with issues such as hand-printed character and cursive handwritten word
recognition. It describes recent achievements, difficulties, successes and challenges in
all aspects of handwriting recognition. It also presents a new approach which
dramatically improves current handwriting recognition systems. Some experimental
results are included.

Introduction

The off-line handwriting recognition problem has been addressed by


many researchers for a substantial amount of time1,2,3,4. Although
isolated character recognition is on its way to being solved, producing
excellent recognition rates1,5, researchers concentrating on the
recognition of handwritten words cannot boast the same success. Two
main approaches for the aforementioned problem have been identified:
1) a global approach2,3,6,7,8 and 2) a segmentation approach9,10,11,12-15.
The first approach entails the recognition of the whole word by the use of
identifying features. The second approach requires that the word be first
segmented into letters. The letters are then recognised individually and
can be used to match up against particular words.
Recently many researchers have been driven to develop off-line
recognition systems due to the challenging scientific nature of the
problem and secondly its industrial importance3,4. The latter arises from
numerous applications of handwriting recognition systems. Some of these
include: postal address recognition, reading machines for the blind,
processing manually filled-out tax forms, and bank cheque
recognition14,16-20.
The remainder of the paper contains 4 sections. Section 2 briefly
describes the history of handwriting recognition systems, Section 3
reviews some recent achievements, a proposed technique is featured in

Section 4, experimental results are presented in Section 5, finally a


conclusion is drawn in Section 6.
2

History of Handwriting Recognition Systems

The history of handwriting recognition systems is not complete


without mentioning the Optical Character Recognition (OCR) systems
which preceded them. Optical Character Recognition (OCR) is a problem
recognised as being as old as the computer itself1,3. There have been
many papers and technical reports published reviewing the history of
OCR technologies21-23. Modern OCR was said to have begun in 1951 due
to an invention by M. Sheppard called GISMO, a robot reader-writer21.
In 1954, a prototype machine developed by J. Rainbow was used to read
uppercase typewritten letters at very slow speeds. By 1967, companies
such as IBM finally marketed OCR systems. However in the late 60s,
these systems were still very expensive, and therefore could only be used
by large companies and government agencies21. Today, OCR systems are
less expensive and can recognise more fonts than ever before. Even so it
is important to note that in some situations these commercial packages
are not always satisfactory. Senior2 mentions that problems still exist
with unusual character sets, fonts and with documents of poor quality.
Research now focusses more on hand-printed numeral, character
and joined/cursive handwriting recognition. Unfortunately the success of
OCR could not carry on to handwriting recognition, due to the variability
in peoples handwriting2. As for the recognition of isolated handwritten
numerals, Suen3, details many researchers which have already obtained
very promising results using various classification methods. Suen
mentions that the key to high recognition rates is feature extraction.
However, this in itself is a very difficult problem which has led
researchers to use more complex methods for preprocessing, feature
extraction and classification. Such methods include the use of Neural
Networks and Mathematical Morphology.
3

Recent Achievements

Researchers all over the world have achieved successful results in


handwriting recognition3,5,24,25. We present some of these results below
in Table 1. As can be seen the table is divided into 3 main categories:
handwritten numeral recognition, character and cursive word
recognition.

Numeral

Character

Cursive
Word

Table 1. Summary of recognition rates


Authors
Recognition Rate [%]
Denker et al. [28]
86
Bottou et al. [27]
91.9
Srihari [19]
89-93
Lee [24]
99.5
Koutsougeras and Jameel [38]
65-98
Liou and Yang [36]
88-95
Srihari [19]
85-93
Shustorovich [37]
89.40-96.44
Edelman et al [33]
50
Lecolinet et al. [2]
53
Leroux et al. [2]
62
Chen et al. [34]
64.9-72.3
Bozinovic et al. [12]
54-72
Senior and Fallside [35]
78.6
Simon [2]
86
Guillevic and Suen [6]
76-98.5
Bunke et al. [26]
98.27

In Numeral recognition there are many excellent results, one of the


best obtained by Lee24. Also in Cursive word recognition the same high
standard of results may be found, Bunke et al.26 obtaining the highest
recognition accuracy. Unfortunately, as previous reviewers have
mentioned, there are too many factors which do not allow a suitable
comparison to be performed. The influencing factors relate directly to the
difference in conditions for experimentation. One of the main differences
is the type of handwriting database used for experimentation. In some
cases, researchers have constrained their experiments heavily, only using
one persons handwriting, while other researchers experiments were not
performed on benchmark databases.
Some of the problems and challenges which are faced by
researchers today include: developing accurate segmentation,
preprocessing, feature extraction and classification techniques. For the
first problem, segmentation, the diverse styles and sizes of handwriting
both play a large factor in the failure of current techniques. In some
cases even a human being would not be able to segment handwriting
containing characters which are tightly packed together and illegible.
These segmentation systems also have to deal with the variability of
handwriting from one person to another, not to mention problems when

one writers handwriting is cursive, while another persons is simply


overlapping.
Challenges faced for preprocessing deal with the choice of whether
to convert raw handwriting into a more efficient form i.e. whether to
binarise the handwriting or keep it in grey-scale form. Other researchers
are disputing whether the handwriting should be thinned or should
remain the way it is to preserve features. Feature extraction poses the
problem of choosing the right features to extract and the right technique
to perform the task. For example researchers may choose between
extracting features such as the entire contours of characters or by
extracting many features such as end-points, loops, holes and so on.
Finally, the task of finding a suitable classification technique (for
individual characters and whole words) has been exhaustively pursued.
However, again the variability of handwriting and the lack of reliable
feature extraction and preprocessing techniques has impaired many
unconstrained approaches. For most of the aforementioned problems,
including feature extraction and classification, researchers have turned
to complex and intelligent methods. The use of Neural Networks has
become extremely popular, and may hopefully enable researchers to
move closer to solving the handwriting recognition problem.
4

Proposed Technique

Following the review of many handwriting recognition systems, this


section contains an explanation of our proposed technique for the
problem of handwriting recognition. Section 4.1, briefly explains the
segmentation system used for separating cursive and joined handwriting.
Section 4.2, explains the use of a Neural Network based dictionary for the
recognition of words.
4.1

Segmentation Technique

The segmentation technique contained two components. Firstly a simple


heuristic segmentation algorithm29 was implemented which scanned
handwritten words for important features to identify valid segmentation
points between characters. The algorithm first scanned the word looking
for minimas or arcs between letters, common in handwritten cursive
script. In many cases these arcs are the ideal segmentation points,
however in the case of letters such as a, u and o, an erroneous
segmentation point could be identified. Therefore the algorithm

incorporated a hole seeking component which attempted to prevent


invalid segmentation points from being found.
If an arc was found, the algorithm checked to see whether it had
not segmented a letter in half, by checking for a hole. Holes, are found
in letters which are totally or partially closed such as an a, c and so
on. If such a letter was found then segmentation at that point did not
occur. Finally, the algorithm performed a final check to see if one
segmentation point was not too close to another. This was done by
ascertaining if the distance between the last segmentation point and the
position being checked was equal to or greater than the average character
width of a particular word. If the segmentation point in question was too
close to the previous one, segmentation was aborted. Conversely, if the
distance between the position being checked and the last segmentation
point was greater than the average character width, a segmentation
point was forced.
The second component of the segmentation technique
incorporated a feedforward artificial neural network trained with the
backpropagation algorithm29. It was initially trained with segmentation
points found through manual segmentation of handwritten words.
Following training, the ANN was presented with segmentation points
obtained through the use of the heuristic segmentation algorithm. The
ANN verified whether all the segmentation points found were correct or
incorrect. Correctly identified points were not removed while incorrect
segmentation points were rejected.
4.2

Neural Network Based Dictionary

After our segmentation technique created a set of segregated characters,


another Neural Network was used to classify the characters. After
classification, these characters were then presented to a neural based
dictionary of words. The network used is based on the Hamming network.
Its architecture includes one input and one output layer which are both
fully interconnected. The input layer accepts ASCII values (divided by
100) of recognised characters which together comprise full words. Each
neuron in the output layer points to a word stored in the dictionary. To
get the desired output, for all i (number of inputs), we subtract the jth
weight (wij) from its corresponding input (xi). For all i we find the total
absolute values of x-w Eq. (1). The output neuron with the smallest value
assigned to it wins Eq. (2) and its corresponding word is chosen as the
correct word. If the case arises that for a particular value of i, |x-w|
equals zero, then a value of -2 is assigned for that particular instance.
Figure 1 shows the network with an example word and dictionary.

x i w ij

(1)

output = min( sum j )

(2)

su m j =

i,j

Figure 1. Hamming network architecture


5

Experimental Results

The new approach has been implemented in C on the SP2 Supercomputer


and some experiments have been conducted. Around fifty handwritten
words were segmented (using the technique detailed in Section 4.1).
Some word samples are presented in Figure 2. The resulting characters
were used to test an ANN trained with a large number of segmented
characters. A feedforward ANN was used, trained with the error
backpropagation algorithm. The ANN architecture consisted of: 700
inputs (character matrix size: 25x28), 26 outputs, 25 hidden units, and a
learning rate and momentum of 0.2. The ANN was trained for 500
iterations.
The identified characters were then presented to the neural
dictionary. The results are provided in Table 2.

Figure 2. Word samples used for experimentation


Table 2. Results using proposed technique
Technique
# of Experiment
Recognition Rate [%]
Character
Word
Proposed approach
Experiment 1
78
62
Without ANN
Experiment 2
66
43
based dictionary
(2 writers)
Proposed approach
Experiment 1
78
100
with ANN based
Experiment 2
66
96
dictionary
(2 writers)
6

Conclusions

We have reviewed many techniques for handwriting recognition,


including numerals and cursive handwritten words. We have also
proposed a technique to improve on current systems. Our results are very
promising indicating that a neural based dictionary can produce
recognition rates of up to 100% for handwritten words. Our results are
comparably better than those of other researchers, however as is the case
with some other researchers, our results are based on a small database of
handwritten words. In future experiments we hope to use a larger
database.
7

References

1.
J-C. Simon, 1992, Off-Line Cursive Word Recognition, Proceedings
of the IEEE, 80, 1150-1161.
2.
A. W. Senior, 1994, Off-Line Cursive Handwriting Recognition
Using NNs, PhD Dissertation, University of Cambridge, England.
3.
C.Y. Suen et al., 1993, Building a New Generation of Handwriting
Recognition Systems, Pattern Recognition Letters, 14, 303-315.

4.
S.N. Srihari, 1992, High-Performance Reading Machines,
Proceedings of the IEEE, 80, 1120-1132.
5.
S-B. Cho, 1997, Neural-Network Classifiers for Recognizing Totally
Unconstrained Handwritten Numerals, IEEE Trans. on Neural
Networks, 8, 43-53.
6.
D. Guillevic, 1997, HMM Word Recognition Engine, ICDAR 97.
7.
D. Guillevic and C. Y. Suen, 1994, Cursive Script Recognition: A
Sentence Level Recognition Scheme, IWFHR94, 216-223
8.
K. Aas, L. Eikvil and T. Andersen, 1995, Text Recognition from
Grey Level Images Using Hidden Markov Models, CAIP95, Prague.
9.
N.W. Strathy, C.Y. Suen, A. Krzyzak, 1993, Segmentation of
Handwritten Digits using Contour Features, ICDAR 93, 577-580
10. G. L. Martin et al., 1993, Integrated Segmentation and Recognition
through Exhaustive Scans or Learned Saccadic Jumps, International
Journal of PR & AI, 7, 187-203.
11. B. Eastwood et al., 1997, A Feature Based Neural Network
Segmenter for Handwritten Words, ICCIMA97, Australia, 286-290.
12. R. M. Bozinovic, S. N. Srihari, 1989, Off-Line Cursive Script Word
Recognition, PAMI, 11, 68-83.
13. H. Fujisawa, Y. Nakano, K. Kurino, 1992, Segmentation Methods
for Character Recognition: from Segmentation to Document Structure
Analysis, Proceedings of the IEEE, 80, 1079-1092.
14. M. Blumenstein, B. Verma, 1997, A Segmentation Algorithm used
in Conjunction with Artificial Neural Networks for the Recognition of
Real-World Postal Addresses, ICCIMA97, Australia, 155-160.
15. S-W. Lee, D-J. Lee, H-S. Park, 1996, A New Methodology for GrayScale Character Segmentation and Recognition, PAMI, 18, 1045-1051.
16. L. Lam et al., 1995, Automatic Processing of Information on
Cheques, International Conference on Systems, Man & Cybernetics, 23532358.
17. D. Guillevic and C. Y. Suen, 1995, Cursive Script Recognition
applied to the Processing of Bank Cheques, ICDAR 95, 11-16.
18. N.W. Strathy, C.Y. Suen, 1995, A New System for Reading
Handwritten Zip Codes, ICDAR 95, 74-77
19. S. N. Srihari, 1993, Recognition of Handwritten and Machineprinted Text for Postal Address Interpretation, Pattern Recognition
Letters, 14, 291-302.
20. M. Gilloux, 1993, Research into the New Generation of Character
and Mailing Address Recognition Systems at the French Post Office
Research Center, Pattern Recognition Letters, 14, 267-276.

21. S. N. Srihari and S. W. Lam, 1995, Character Recognition,


Technical Report, CEDAR-TR-95-1.
22. V.K. Govindan, 1990, Character Recognition - A Review, Pattern
Recognition, 23, 671-683.
23. S. Impevedo et al., 1991, Optical Recognition - A Survey,
International Journal of PR & AI, 5, 1-24.
24. S-W. Lee, 1996, Off-Line Recognition of Totally Unconstrained
Handwritten Numerals Using MCNN, PAMI, 18, 648-652.
25. S-W. Lee, 1995, Multilayer Cluster Neural Network for Totally
Unconstrained Handwritten Numeral Recognition, NNs, 8, 783-792.
26. H. Bunke et al., 1994, Off-line Cursive Handwriting Recognition
Using HMMs, Technical Report IAM-94-008, University of Bern.
27. L. Bottou et al., 1994, A Comparison of Classifier Methods: A Case
Study in Handwritten Digit Recognition, Proc. IEEE Workshop on
Applications of Computer Vision, Sarasota, Florida, 77-82.
28. J. S. Denker et al., 1989, Neural Network Recognizer for
Handwritten Zip Code Digits, NIPS, 1, 323-331.
29. M. Blumenstein and B. Verma, 1998, An Artificial Neural Network
Based Segmentation Algorithm For Off-Line Handwriting Recognition,
International Conference on Computational Intelligence and Multimedia
Applications (ICCIMA 98), 27-33.
30. T. M. Breuel, 1993, Recognition of Handprinted Digits Using
Optimal Bounded Error Matching, Technical Report #93-06, IDIAP,
Switzerland.
31. T.M Breuel, 1994, A System for Off-Line Recognition of
Handwritten Text, Technical Report #94-02, IDIAP, Switzerland.
32. S-W. Lee, D-J. Lee, H-S. Park, 1996, A New Methodology for GrayScale Character Segmentation and Recognition, IEEE Transactions on
Pattern Analysis and Machine Intelligence, 18, 1045-1051.
33. S. Edelman, T. Flash and S. Ullman, 1990, Reading Cursive
Handwriting by Alignment of Letter Prototypes, International Journal of
Computer Vision, 5, 303-331.
34. M-Y. Chen, A. Kundu and J. Zhou, 1994, Off-line Handwritten
Word Recognition using a Hidden Markov Model Type Stochastic
Network, IEEE Transactions on Pattern Analysis and Machine
Intelligence, 16, 481-496.
35. A. W. Senior and F. Fallside, 1993, Using Constrained Snakes for
Feature Spotting in Off-Line Cursive Script, International Conference on
Document Analysis and Recognition (ICDAR 93), 305-310.

36. C-Y. Liou and H-C. Yang, 1996, Handprinted Character Recognition
based on Spatial Topology Distance Measurement, IEEE Transactions on
Pattern Analysis and Machine Intelligence, 18, 941-944.
37. A. Shustorovich, 1994, A Subspace Projection Approach to Feature
Extraction: The Two-Dimensional Gabor Transform for Character
Recognition, Neural Networks, 7, 1293-1301.
38. C. Koutsougeras and A. Jameel, 1995, Experiments with Various
Architectures for Handwritten Character Recognition, IEEE Proceedings
of the 37th Midwest Symposium on Circuits and Systems, 573-576.

S-ar putea să vă placă și